Research On Key Technologies Of Fixed-point Monitoring Of Illegal Buildings

The illegal exploitation of land and illegal construction seriously affect social harmony and sustainable development,and the "early detection,early investigation and punishment" of land violations is very important in the process of urbanization.In order to solve the problem of low manual inspection efficiency and low camera monitoring intelligence,this paper focuses on the key technologies in the automatic recognition of illegal buildings,including fast acquisition of monitoring point image,change detection of illegal buildings,vehicle false alarm filtering.The main research contents are as follows:(1)A virtual preset position setting method for fixed-point monitoring based on pyramid L-K optical flow method is proposed.In order to solve the problem that the camera has insufficient preset positions,firstly,gray-scale transformation and median filtering are used to remove the noise points in the two adjacent frames in the video.Secondly,pyramid-based Lucas Kanade optical flow algorithm is used to calculate the optical flow of the strong corner points in the previous frame,and the inter-frame motion vector is obtained by calculating the affine transformation.Then the motion vector of adjacent frames is integrated to get the total motion vector of each frame.Finally,binary search is used to extract the required virtual preset frame,which has the closest motion vector in the video.Experiments show that the overlap rate between the virtual preset position and the reference image is more than 95% and the algorithm is resistant to light and weather changes.(2)A new and old phase monitoring image change detection algorithm based on level set model is researched.To solve the problem that the difference method is easy to lose the information of the effective change area,firstly,the old and new phase monitoring images are matched and the neighborhood information is used to convert it to the average grayscale image.Then the difference map is obtained by applying the logarithmic ratio method to suppress the non-changing area information and normalization operation.Finally,the level set model is used to segment the difference map to extract the change area.This method has an average Kappa coefficient of 0.83 and 0.81 on the measured data of Anshun and Liuying respectively.Experiment prove that this method has more continuously changing areas segmentation,and can effectively reduce the pseudo-changing pixels.(3)A small vehicle detection algorithm based on improved Faster R-CNN is researched.In order to solve the problem that vehicles interfere with the detection of illegal buildings and the high miss rate of small target vehicles in Faster R-CNN algorithm,ML-FPN is used to extract multi-level and multi-scale feature pyramids and integrate them into Faster R-CNN framework.To solve the mismatch between ROI and feature map,ROI Align is used instead of ROI Pool.Experiments show that the average detection accuracy of vehicles has increased from83.6% to 93.8%,effectively improving the detection capabilities of small target vehicles.An integrated experiment was carried out on the automatic detection algorithm of illegal buildings in this article.The highest detection accuracy in Anshun city Guizhou province area reaches 91.1% and the missed detection rate was only 6.7%;the highest detection accuracy in Liuying city Anhui Province area reaches 93.4% and the missed detection rate was only 7.8 %.